rncz.net

ComfyUI on Windows 11 with ROCm (Official 2026 Installation): Z-Image Turbo on AMD GPUs

comfyuirocmamdpytorchz-imageimage-generationwindows

If you have an AMD GPU with the RDNA 3 or RDNA 4 architecture, this is the guide you need: install ComfyUI on your Windows PC with ROCm — now using the official installation instructions from the ComfyUI repo — and generate images with Z-Image Turbo. This works on the RX 7000, RX 8000, and RX 9000 series as well as Strix Halo (RDNA 3.5).

The good news: the ComfyUI GitHub now has a dedicated AMD GPUs → Windows section with a ready-made pip command, so you no longer need to dig through the ROCm repo yourself.

This post is a continuation of the Flux.1-dev and Wan 2.2 articles — but you can also use it as a fresh install guide.

1. Install Git SCM

Head over to git-scm.com, click the button, and download the Git for Windows 64-bit Setup. Run the executable — the defaults are fine.

Verify the installation in a terminal (Win key, type cmd, Enter):

git --version

If a version is printed, you’re good.

2. Install Python 3.13

Go to python.org/downloads, grab the Windows installer for Python 3.13, and run it (tick Add Python to the PATH).

Open a new command prompt and check:

python --version

If you have multiple Python versions installed, use the py launcher to target the one you want:

py -3.13 --version

3. Clone ComfyUI

Head to the ComfyUI GitHub repository and scroll down to the AMD section. You’ll see separate instructions for Linux and Windows — we use the Windows one.

The page lists the supported GPUs:

  • RX 7000 series — RDNA 3
  • RX 9000 series — RDNA 4
  • Ryzen AI Max (Strix Halo) — RDNA 3.5

Back at the top of the page, hit the Code button, copy the clone URL, and clone the repo:

mkdir AI
cd AI
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

To update ComfyUI later, just run git pull. If something ever breaks, you can always delete the folder and clone again.

4. Create and activate a virtual environment

With the repo checked out, create a virtual environment with Python 3.13:

python -m venv venv

(or py -3.13 -m venv venv if you have multiple versions installed), then activate it:

venv\Scripts\activate

5. Install PyTorch with ROCm

Back in the ComfyUI repo, go to the AMD / Windows section again — there’s a ready-made command to install Torch, TorchAudio, and TorchVision with ROCm support. Copy it and paste it into your (activated) terminal:

pip install "torch[device-all]" "torchvision[device-all]" "torchaudio[device-all]"

Wait a moment while it downloads. You’ll see it install wheels for the gfx1151, gfx1152, and gfx1153 variants. If you only want the wheel for your GPU (smaller download), change device-all to device-gfxXXXX — for example:

pip install "torch[device-gfx1152]" "torchvision[device-gfx1152]" "torchaudio[device-gfx1152]"

Verify everything landed:

pip list

You should see the ROCm packages together with torch, torchaudio, and torchvision.

6. Install ComfyUI’s requirements

pip install -r requirements.txt

7. Run ComfyUI

Start the server:

python main.py --highvram

The --highvram flag makes ComfyUI use as much of your VRAM as possible for the loaded models.

In the console you’ll see that ComfyUI detects your AMD card (e.g. an RX 7900 XTX with gfx1100 and 24 GB of VRAM) — no NVIDIA card required. Grab the URL it prints (usually http://127.0.0.1:8188) and open it in your browser.

8. Set up Z-Image Turbo

In the ComfyUI UI, open the menu at the top left → Browse templates → Z-Image Turbo Text to Image.

You’ll see red errors — the workflow needs three models that aren’t downloaded yet. Download them from Hugging Face and put each one in the right directory:

ModelDirectory
z_image_turbo_bf16.safetensorsmodels/diffusion_models
ae.safetensorsmodels/vae
Quantized ~4B parameter text encodermodels/text_encoders

Once the files are in place, hit refresh in the browser — the errors are gone.

9. Generate your first image

Type a prompt — the classic one from the video is “a pelican on a bicycle, cartoon” (a prompt normally used to test coding models, now feeding an image model instead). You can also enable the console in the UI to watch everything load.

Queue the workflow. On an RX 7900 XTX you get around 1 iteration per second — pretty nice for a local model — and after a short wait, there’s your image, generated entirely on your local AMD GPU.

Next steps

  • Want more? ComfyUI works with far more models now that the AMD/ROCm stack is set up — see the Flux.1-dev and Wan 2.2 video articles.
← All posts